X-Authentication-Warning:	delorie.com: mail set sender to djgpp-bounces using -f
Date:	Wed, 17 Nov 2004 15:56:52 +0500
From:	Alexander Mironenko <a_mironenko AT mail DOT ru>
X-Mailer:	The Bat! (v1.00 Build 1336) UNREG
Message-ID:	<18664.041117@mail.ru>
To:	djgpp AT delorie DOT com
Subject:	Re[2]: msoft-float option with DJGPP
References:	<200411160505 DOT iAG551Qb009789 AT delorie DOT com>
Mime-Version:	1.0
X-OriginalArrivalTime:	17 Nov 2004 10:58:53.0966 (UTC) FILETIME=[6DA03AE0:01C4CC94]
Reply-To:	Alexander Mironenko <a_mironenko AT mail DOT ru>

>> I know about "-lemu" option but in my case this option means severe
>> slowdown (up to several times).

> And what makes you think -msoft-float wouldn't cause that same kind
> of slowdown?

I have a Borland C application (let's say a game) giving me 20 fps
on FPU-less board. It was compiled with Borland's emulation option.

I replaced graphic output routines with allegro's (game programming
library) analogs and recompiled game with DJGPP and "-lemu" option.
Result: 7 fps.
It seemed to me very strange, since before I have tested allegro's
speed on same amount of random graphic output and had had more than 70 fps.

So I digged into my source and found that every frame I do
approximately 400 floating point multiplications. I replaced
most of them with fixed point analogs (all numbers were relatively
small) and now instead of 7 fps I have 57 fps - it means that
graphic routine was fast (I haven't changed it) and floating point
calculations were slow.

That's why I hope that integrated at compliling time emulation
(like Borland's one - more than 20 fps)
will be significally faster than (interrupt driven) libemu (7 fps).

That's why I have asked about "-msoft-float" support
in next release of DJGPP's gcc, since now
it is relatively easy - IFAIK such support is incorporated in
modern gcc versions.
Please correct me if I am wrong.

--
Alexander Mironenko

- Raw text -